Multi-Agent Risks from Advanced AI
Advanced AI systems are increasingly interacting with each other autonomously - from trading bots over swarms of research agents to fleets of autonomous vehicles.
This multi-agent reality is arriving faster than our understanding of its implications. Drawing on the recent "Multi-Agent Risks from Advanced AI" report, this talk examines the unique challenges emerging from AI systems interacting at scale.
BuzzRobot guest Jason Hoelscher-Obermaier, Director of Research at Apart Research (a nonprofit AI safety research lab), will introduce three key failure modes unique to multi-agent systems:
- miscoordination (where AIs with shared goals still fail)
- conflict (where resource competition can rapidly escalate)
- collusion (where AIs cooperate against human interests).
Jason will illustrate these risks using case studies that give a taste of both theoretical concerns and observable impacts, drawing from both real-world deployments and simulations. These examples demonstrate emerging problems like algorithmic price coordination, failures in critical coordination tasks despite aligned goals, and how AI systems might collaborate to evade safety mechanisms.